Best AI Voice Generators 2026: 5 Tested, 10 More Compared

An AI voice-over generator turns a written script into narration. You paste text, pick a voice, and get an audio file for a YouTube explainer, a course lesson, a product demo or an ad. There are dozens of these tools, and most of them sound good on their own demo lines. They differ in the places a demo hides: how they read prices, dates, names and acronyms, how loud the files come out, what the free plan actually lets you do, and what a million characters costs.

So instead of ranking tools from their marketing, I ran the same two short scripts through five engines and measured the results, then compared ten more from their own pricing pages and docs.

Quick picks

• Free, offline, and fine for commercial use: Kokoro-82M. Apache-2.0 licence, passed all 10 traps in the first script, but it comes out quiet, so normalise it.
• Try before you sign up: Murf’s and OpenAI’s web demos need no account. Paste up to 500 characters of your own script into Murf’s, or a short script into openai.fm, and listen.
• The tool most reviews put first: ElevenLabs. Commercial use starts at $6 a month. I couldn’t test it.
• English e-learning with unlimited re-takes: WellSaid bills only the minutes you download. I couldn’t test it.
• Developers on a budget: Google and Amazon’s basic voices cost $4 per million characters; Azure’s neural voices cost $15, with 500,000 free characters a month (table).
• No subscription: Narakeet sells 30 minutes for a one-time $6.

How I tested

Two scripts, each about 25 seconds long, written to trip a voice up. Every engine got the same text, pasted as written, with its default settings.

For each take I measured:

• Pace in words per minute (WPM), over the spoken part only.
• Loudness in LUFS and true peak in dBTP, on an EBU R128 meter. Apple’s podcast spec, for example, asks for about −16 LUFS ±1 dB with true peaks no higher than −1 dB (Apple). A true peak above 0 means the file can clip.
• Pauses of 0.25 s or longer.
• The 20 traps, pass or fail. I used a speech-recognition transcript for every take. For the three takes hosted here, the phonemes the open models produced and a second AI listening model were extra checks. The two demo takes were checked from the transcript alone, so their intonation and heteronym items are marked “not checked”.

The players on this page are loudness-matched to −16 LUFS, so a louder file doesn’t sound “better”. The results panel shows each file’s raw loudness as it came out of the tool.

Results

What stands out:

• Only three of five read “$4.99” as a price. Kokoro and OpenAI said “four dollars and ninety-nine cents”, and Murf said “four dollars ninety-nine cents”. The other two read it as a decimal: “four point nine nine dollars” and “dollar four point nine nine”.
• Loudness varies by 9 dB before you touch anything. Kokoro came out at about −26 LUFS, so it needs about 10 dB of gain to sit with the rest. Piper went the other way: its true peak is slightly over 0 dBTP, so it can clip after MP3 conversion unless you limit it.
• The two commercial demos paused the most and read the slowest (144–150 WPM, 10–14 pauses). The other three ran at 155–170 WPM with 4–9 pauses. Neither is wrong: slower, more paused reading suits an explainer, and faster suits a walkthrough.
• Speech ran at about 760–900 characters a minute. That matters when a plan sells minutes instead of characters (prices).

Here are the three takes I can host. Murf’s free plan includes no downloads or commercial rights, and I couldn’t load OpenAI’s terms page to confirm what it allows, so for those two, paste the scripts into their demos and listen there.

The trap scorecard

The 20 traps are listed under the two scripts above. Each engine below shows only what it missed, what was unclear and what I couldn’t check; it passed everything else.

Three caveats. The transcript wrote “March three” for every voice, even where the phonemes and the AI listener heard “third”, so it can’t settle that one. The AI listener can be wrong too: where it disagreed with the phonemes, I marked the trap unclear. And the Murf and OpenAI results come from the transcript alone, so “10 of 10” there means every item a transcript can check, not every trap.

Fixing the misreads

Every engine has a way to force a pronunciation. I tried each tool’s own method on the words it got wrong:

• Kokoro accepts inline phonemes: [SAML](/sˈæməl/). Its phoneme output shows both SAML and “Joaquín” changed to the right sounds.
• Piper accepts raw phonemes in double brackets: [[ fˈɔːɹ nˈaɪnti nˈaɪn ]]. That fixed the price (the transcript now says “four ninety-nine”), but “aloud” in the sentence before came out blurred: the transcript heard “allel”.
• SSML APIs take <phoneme> tags. In my API test the tag brought “Joaquín” closer, but on the price it went wrong (below). Typing “four ninety-nine” into the script worked.

Two lessons that apply to any tool:

• Respelling beats markup. Writing the word the way it’s said (“four ninety-nine”, “sam-ul”) works everywhere and can’t break the file. Keep phoneme tags for names you can’t respell.
• A fix can change the whole take. The “four ninety-nine” take above finished in 23.0 s instead of 27.1 s, and its six pauses of 0.25 s or more dropped to none, though only the price changed. Listen to the full file again after any fix, not just the word you fixed.

Prices per 1M characters

Plans are sold in characters, credits, minutes or hours, which makes them hard to compare. The list below converts each plan to dollars per million characters, about 20 hours of narration. Where a plan sells minutes, I converted at 850 characters a minute, the middle of what the test voices actually spoke. The vendors don’t state these per-million figures; I worked them out from the prices below. All prices are in USD, as listed on Oct 20, 2026.

Read the “Free audio, commercial use” line before you build on a free plan. Of the hosted tools whose terms I checked, only Typecast lets free audio be used in public work, and only with attribution. The open-source models are the real free option.

The generators, one by one

The five I tested come first, then the ten I could only read about. Each “from the docs” entry says so.

Kokoro-82M: the best free option I tested

What it is: an open-weight model with 82 million parameters that ran fine on an ordinary CPU. Its model card lists the Apache-2.0 licence, 54 voices in 8 languages for v1.0, and training “exclusively on permissive/non-copyrighted audio data”.

• What it got right: all 10 Script A traps, including the date, “Siobhan”, both readings of “read”, a rising question, and the price as “four dollars and ninety-nine cents”.
• What it got wrong: it spelled SAML out letter by letter, said “Joaquín” with an English J, and may have said “live” to rhyme with “give” (the phonemes say “lyve”; the AI listener heard “liv”).
• Watch out for: loudness. At −25.8 LUFS it was 3 dB quieter than the next-quietest take and about 10 dB below the −16 LUFS the players here use. Normalise it before you publish.
• Speed: about 80–90 seconds to render a 27-second clip on a CPU, with no GPU.
• Price: free to run yourself. The model card cites hosted APIs at under $1 per million characters.
• Best for: developers and tinkerers who want free, commercial-safe narration and don’t mind a command line. Skip if you want a point-and-click editor.

Murf: tested through its free web demo

What it is: a voice-over studio aimed at business and marketing work, with a TTS API. Its text-to-speech page has a free demo box: “Free to use. 35 languages. 200+ lifelike voices. No sign-up required.” It takes up to 500 characters.

• Measured: the slowest and most-paused reads of the test (146 and 144 WPM, 11 and 14 pauses), at −20 LUFS. Audio started about 10 seconds after I pressed play.
• Traps: all 10 items a transcript can check came out right, including “four dollars ninety-nine cents”, “example dot com” and SSO, CRM and CSV as letters. Names, SAML, intonation and the heteronyms need a listener, so they’re unchecked.
• Price: Creator is $19 a month billed yearly ($228) for 24 hours of generation a year; Business is $66 a month billed yearly. The free plan gives 10 minutes with “No Downloads” and “No Commercial Rights”. The API costs $0.01 per 1,000 characters for its Falcon model and $0.03 for its studio-quality model, with $10 of free credit (pricing).
• Best for: slower, more paused narration for business videos, and anyone who wants to hear their own script before signing up.

OpenAI gpt-4o-mini-tts: tested through openai.fm

What it is: OpenAI’s text-to-speech API. openai.fm is its free public demo: pick one of 11 voices, optionally describe a “vibe” in plain words, paste a script and press play. I used Marin with no vibe.

• Measured: 150 and 146 WPM, 10 and 13 pauses, −22.7 LUFS, with plenty of headroom (true peak −3.2 dBTP or lower). The demo returned 24 kHz MP3 about 10 seconds after I pressed play.
• Traps: all 10 items a transcript can check came out right, including “four dollars and ninety-nine cents” and “example dot com”. Names, SAML, intonation and the heteronyms are unchecked.
• Price: the API bills gpt-4o-mini-tts by token ($0.60 per million text tokens in, $12 per million audio tokens out); the older tts-1 is $15 per million characters and tts-1-hd $30 (pricing). OpenAI requires you to tell listeners the voice is AI-generated.
• Best for: developers who want to steer delivery with plain-English instructions. Skip if you want an editor, not an API.

Piper: the fastest, with a licence catch

What it is: a fast, lightweight open-source TTS engine that runs offline. The current package is piper1-gpl (version 1.8.0, GPL-3.0), and its README says it’s “Looking for Maintainers”.

• What it got right: the date, the year, “1,250”, the heteronyms and every Script B acronym.
• What it got wrong: “dollar four point nine nine”, a flat question and “Joaquín”. It also treated “Dr.” as the end of a sentence and paused after it. “Siobhan” is unclear: its phonemes say “shi-VAWN”, but the AI listener heard “chev-on”.
• Watch out for: clipping (true peak +0.1 dBTP), the lowest sample rate in the test (22.05 kHz), and voice licences. The code is GPL, but each voice has its own licence. The lessac voice I used was trained on a dataset licensed “for Research Purposes only”, which rules out commercial voice-synthesis products. Check the model card of any Piper voice before you use it commercially.
• Speed: about 13–15 seconds per clip on a CPU, the fastest render in the test.
• Best for: offline apps, prototypes and accessibility tools, where speed matters more than polish.

Lazybird (ours): the lowest studio price per character here, with bugs

What it is: our browser studio for voice-overs and narrated videos, with an API and voice cloning. Its voice list returned 560 voices across 151 locales when I checked.

• What it got right: the date, the year, “Dr.”, “Siobhan”, “1,250”, the heteronyms, and every Script B acronym and number, with no glitches in the raw takes.
• What it got wrong: all four voices I tried (Ava, Andrew, Brian, Emma) said “four point nine nine dollars”. Ava’s question ended flat, and “Joaquín” came out as roughly “shwah-kin”.
• Fixes: typing “four ninety-nine” fixed the price but sped up the whole take. A one-word phoneme tag brought “Joaquín” closer (“wah-kin”), though the vowel was still short.
• Bugs I found: a multi-word phoneme tag on the price sounded robotic and added about 3 seconds of garbled speech after the last sentence. It happened with two voices out of two. The API accepts only <break> and <phoneme>; a <sub> tag returned an error. Plain-text renders took 13–27 seconds; SSML and edited renders took 43–94 seconds.
• Measured: 155 and 170 WPM at about −18 LUFS, the closest to a publishable level out of the box, as 48 kHz MP3.
• Price: one credit is one character. Free: 10,000 credits a month, personal use only. Creator $49 a year (500,000 credits a month, commercial rights, cloning), Pro $99 a year (2M), Max $199 a year (10M); monthly billing costs more (pricing).
• Best for: long narration on a budget, and turning a script into a narrated video in one place. Skip if you need the most expressive delivery.

ElevenLabs (from the docs)

The tool most reviews put first (Zapier, Nerdynav and G2 all lead with it). Its pricing page now opens with “Introducing Eleven v4”.

• Price: Free 10,000 credits a month; Starter $6 (30,000 credits, “Commercial License”, instant cloning); Creator $22 (121,000 credits, professional cloning; first month 50% off); Pro $99 (600,000 credits, 192 kbps MP3 and 44.1 kHz PCM through the API). For its multilingual models, one character is one credit. API: v4 at $0.08 per 1,000 characters, “90+ languages”; Flash and Turbo at $0.04 (pricing).
• Free plan: its terms say free users “may only use the Services for non-commercial purposes”.
• Not tested: the homepage demo blocked my attempts.

WellSaid (from the docs)

• Price: Starter $10 a month billed yearly ($120 for 240 downloaded minutes a year) or $19 monthly for 20 minutes; Pro $33 a month billed yearly (2,160 minutes a year). Generation is unlimited; you pay for what you download (pricing).
• Good to know: individual plans get “All English voices” only; Starter exports 24 kHz MP3. All paid plans include commercial rights; the free trial has none.
• Not tested: the homepage demo stops at 200 characters.

Fish Audio (from the docs)

• Price: Free 8,000 credits a month (about 7 minutes, 500 characters per take); Plus $11 a month billed yearly or $15 monthly (about 200 minutes, commercial use); Pro $75 or $100 (plans). The API costs $15 per million UTF-8 bytes, roughly a million characters of English (docs).
• Good to know: its FAQ says a minute costs “roughly 600 to 625 credits”, but the plan cards work out to about 1,140–1,250, so trust the minutes on the cards. A pop-up on Oct 20 offered 50% off all plans until November 12; the prices above are list prices. Its older S1 model retires on Dec 31, 2026.
• Not tested: the homepage demo’s reCAPTCHA check failed in my browser.

Cartesia (from the docs)

• Price: Free 20,000 credits a month; Pro $5 (100,000 credits, commercial licence, instant cloning); Startup $49 (1.25M); Scale $299 (8M). “1 credit equals 1 character”, and a minute uses 750–800 credits (pricing).
• Good to know: API-first, aimed at developers. Its terms bar using the output to build competing products.
• Not tested: trying it needs an account.

Typecast (from the docs)

• Price: Free 3,000 credits once (about 5 minutes), and “Attribution is required for all content downloaded on the Free plan”. Basic $5 a month (30,000 credits, about 35 minutes, commercial licence, one cloning slot); Plus $19; Pro $29 (pricing).
• Not tested: the homepage demo takes 300 characters and offers emotion presets, but neither script fits.

Descript (from the docs)

• What it is: a video and podcast editor first, where AI speech fixes or adds lines in a recording. It’s a different job from a voice-over generator.
• Price: Hobbyist $16 a month billed yearly or $24 monthly (400 AI credits a month, “AI Speech with custom voice clones”) (pricing). Credits aren’t characters, so it has no per-million figure.
• Not tested: needs an account.

Narakeet (from the docs)

• Price: no subscription. “The 6 USD package gives you 30 minutes”, and bigger packs cost less per minute. The free tier gives 20 conversions of scripts up to 1 KB, with “Commercial use: No” (pricing).
• Good to know: it also turns slide decks into narrated videos. Its terms forbid republishing material from its site, so I can’t host its demo audio here.
• Not tested: needs an account.

Left out on purpose

• Hume AI (Octave): its terms (§4.2) forbid customers to “publish benchmarks or performance information about the Service”, so I didn’t measure it.
• PlayHT: Meta acquired PlayAI (PlayHT) in July 2025 (TechCrunch), and play.ht no longer resolves.
• LOVO (Genny): lovo.ai returned an HTTP 402 “deployment disabled” error on Oct 20, 2026, so its website was offline when I tested.

For developers: cloud TTS APIs

If you’re generating speech from code, the big clouds are much cheaper per character than the creator studios. These are list prices per million characters, checked Oct 20, 2026.

What’s new as of October 2026

• ElevenLabs is promoting v4 (“Introducing Eleven v4”), at $0.08 per 1,000 characters through the API.
• Murf’s API lists Falcon at $0.01 per 1,000 characters, a third of its Gen2 price.
• Fish Audio is retiring its S1 model on Dec 31, 2026; its homepage demo now says “Powered by Fish Audio Drama 3”.
• PlayHT is gone after Meta’s 2025 acquisition, and LOVO’s site was offline in October.
• Piper’s current home is a GPL-licensed repository that says it is looking for maintainers.

How to choose

1. Check the licence first. If the audio goes into anything public or paid, a free hosted plan almost never allows it. Budget for the cheapest paid plan, or use Kokoro.
2. Paste your own script, not theirs. Use the hardest lines you have: prices, product names, people’s names, acronyms. Murf’s and OpenAI’s demos let you do this with no account.
3. Count characters, not “minutes”. A minute of speech is about 850 characters. Multiply your monthly script length by the per-million price above.
4. Check how you’ll fix a word. Respelling always works. Phoneme markup is better for names, but test it on a short line first.
5. Plan to normalise. Raw files ranged from −26 to −17 LUFS. Whatever you pick, level it to your platform’s target before you publish.

By job:

• YouTube explainers: a slower, pausier voice. Murf’s Natalie was the slowest in the test. Try your script in Murf’s and OpenAI’s demos, then pick on sound.
• Courses and walkthroughs: watch the acronyms. Every engine got SSO, CRM and CSV right, but Kokoro spelled out SAML. WellSaid is built for English training content.
• Apps and automation: Azure, Google or Polly for cost; OpenAI if you want to steer the delivery in words; Kokoro if you want to run it yourself.
• Long narration on a small budget: compare the per-million prices. Plans sold by downloaded minute cost the most for long scripts.

Best ElevenLabs alternatives

• For less money per character: Cartesia ($37–50 per 1M) or Fish Audio ($15 per 1M through its API).
• For free, commercial-safe output: Kokoro-82M.
• For instructions in plain English: OpenAI’s gpt-4o-mini-tts.
• For English training videos: WellSaid.

Test any generator with these two scripts

Copy Script A and Script B from above, paste each into the tool, and listen for these:

1. Does “$4.99” sound like a price? (Two of the five engines failed this.)
2. Does “Dr.” become “Doctor”, without a pause after it?
3. Is it “shi-VAWN” and “hwah-KEEN”, or something closer to the spelling?
4. Is the first “read” “red” and the second “reed”? Is it “leed times” and “lyve”?
5. Does SAML become a word? Do SSO, CRM and CSV stay as letters?
6. Do both questions rise at the end?
7. Is anything added, skipped or repeated, especially at the very end?

To check loudness, run ffmpeg -i take.mp3 -af ebur128 -f null - and read the “I:” value at the end. For podcasts, Apple asks for about −16 LUFS.

FAQ

What is the best AI voice generator?

It depends on the job and the licence. In this test, Kokoro-82M was the strongest free option, and Murf’s and OpenAI’s demos are the quickest way to hear a commercial voice on your own script. ElevenLabs is the most common pick in other reviews, but I couldn’t test it. No listening panel has rated these voices yet, so I won’t call any of them the most realistic.

Is there a free AI voice generator I can use commercially?

Kokoro-82M, under Apache-2.0, run on your own computer. Piper’s code is free, but check each voice’s licence. Of the hosted tools, ElevenLabs, WellSaid, Fish Audio, Cartesia and Narakeet don’t allow commercial use on their free plans, Murf’s free plan has no downloads, and Typecast requires attribution.

How much does an AI voice-over cost per minute?

At about 850 characters a minute: around $0.003 on Google’s WaveNet ($4 per 1M), $0.013 on OpenAI tts-1 ($15), $0.15 on ElevenLabs Creator ($182 per 1M), and $0.50 on WellSaid Starter billed yearly.

How do I make an AI voice pronounce a word correctly?

Respell it in the script (“four ninety-nine”, “sam-ul”). If that isn’t enough, use the tool’s phoneme or pronunciation feature, and listen to the whole take afterwards, because the fix can change the pacing.

Why do AI voices read prices and dates oddly?

The text is converted to words before it’s spoken, and “$4.99” has several valid readings. Two of the five engines read it as a decimal (“four point nine nine dollars”). Writing the words you want removes the guesswork.

What happened to PlayHT?

Meta acquired PlayAI, the company behind PlayHT, in July 2025. play.ht no longer resolves.

Why isn’t Hume AI tested?

Hume’s terms forbid customers from publishing benchmarks or performance information about its service.

Sources and test log

• Test date: Oct 20, 2026. Two scripts, raw take 1 from each engine, default settings, no cherry-picking.
• Engines: Kokoro-82M v1.0 (voice af_heart) and piper-tts 1.8.0 (en_US-lessac) on a CPU; Lazybird API (Ava, plus Andrew, Brian and Emma for the price check); Murf web demo (Natalie, Conversational); openai.fm (Marin, gpt-4o-mini-tts, no vibe).
• Measurement: EBU R128 loudness, silence detection at −35 dB, speech-recognition transcripts, and a second AI listening model for the three hosted takes. Players are loudness-matched to −16 LUFS, 128 kbps MP3.
• Prices and terms (checked Oct 20, 2026): ElevenLabs, Murf, WellSaid, Fish Audio, Cartesia, Typecast, Descript, Narakeet, Kokoro, Piper, and the cloud API pages linked in the developer list.

For recording your own voice instead, see our voice-over recording guide and microphone picks. For free apps, see free voice-over software, and for putting narration on video, text to speech for video and TikTok voiceovers.